Most AI headlines are about the biggest model: the highest score, the longest context, the most expensive training run. The more important story for anyone actually running AI agents is usually at the other end of the lineup. On Monday, Meta Superintelligence Labs released Muse Ember, its lightest and most cost-efficient model so far, and the pitch is almost entirely about price.
Meta describes Ember as designed for “complex reasoning execution and the work agents repeat all day”: tool calls, coding subtasks, routing, extraction and long document parsing. Boring jobs, in other words, but they make up most of what production agents do, and they add up fast on a bill.
The short version
- Muse Ember is priced at 10 cents per million input tokens
- Meta says it beats Anthropic’s Haiku and OpenAI’s Luna on cost per task, using the Artificial Analysis composite benchmark
- It is cost-competitive with open models like DeepSeek V4.1 Flash and GLM 5.3 Flash
- It handles text, images and audio in a single model
- It is meant to pair with the larger Muse Wonder: “Wonder plans, Ember handles the volume”
The fourth Muse model
Ember completes a lineup that Meta has rolled out over about seven months. If you have lost track, here is the order of release.
Spark arrived in April as the mid-weight model, and Glimmer followed in August as its open source twin. We looked at that release in Meta’s quiet 30B open-weight model on Hugging Face. Wonder, the first frontier model, landed in October. Meanwhile the consumer-facing Muse Agent launched in September and climbed the US app charts faster than many expected. We covered the trust question around it in our piece on Meta’s AI agent that can email, shop and pay for you.
How the pricing compares
At 10 cents per million input tokens, Ember sits well below the frontier-adjacent small models from Anthropic and OpenAI, according to Meta’s own comparison. It also lands in the same price zone as the open-weight “flash” models from DeepSeek and Zhipu, which have been the default budget choice for many developers.
To make that number concrete, here is a quick bit of arithmetic. This counts input tokens only, since output pricing was not part of the launch coverage we reviewed, so treat it as a floor rather than a full bill.
| Workload (input only) | Tokens | Cost at $0.10 per million |
|---|---|---|
| 10,000 support tickets at 2,000 tokens each | 20 million | $2.00 |
| 100,000 invoices parsed at 3,000 tokens each | 300 million | $30.00 |
| 1 billion tokens of routing and extraction | 1 billion | $100.00 |
Those are illustrative examples, not Meta’s figures, but they show why a tenth of a dollar per million tokens changes what is worth automating. Jobs that were too expensive to run on every record start to look reasonable.
“Wonder plans, Ember handles the volume”
The most useful idea in the announcement is architectural rather than numerical. Meta designed Ember to sit under a bigger model inside an agent harness. A heavy model like Wonder works out the rules or the plan, then hands the repetitive execution to the cheap one.
Picture a customer-support agent. The large model reads a tricky policy and decides how a particular kind of refund should be handled. Ember then applies that decision across tens of thousands of tickets. You pay frontier prices for the thinking once, and budget prices for the doing many times.
One model for text, images and audio
Ember is also multimodal. Meta says agents can read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, all without a separate model. That matters for the cost story. Many real pipelines used to chain a transcription model, a vision model and a text model. Folding those into one cheap model simplifies the system and shrinks the bill at the same time.
| Input type | Example agent job |
|---|---|
| Text | Classify and route incoming support emails |
| Images | Read a dashboard screenshot or a scanned receipt |
| Audio | Summarize a meeting recording into action items |
| Video | Pull key moments out of a recorded demo |
Cheap models are where the market is heading
Ember is not an isolated release. The industry keeps pushing prices for capable small models down, and each lab wants to own the “workhorse” slot. We saw OpenAI go after that same ground in GPT-6.1 Terra’s $1 per million token pricing, and Nvidia is chasing local, desk-sized agents with its Nemotron 3.5 models squeezed into 24GB. Ember’s headline price is a sharp bid in that race.
Read the claims carefully
- The “cheaper and smarter” comparison comes from Meta’s own announcement, so test it on your workload
- Input price is only part of the bill. Check output token pricing, rate limits and context limits
- Composite benchmark scores can hide weak spots on specific tasks like your document formats
- A cost-per-task win depends on how many retries the cheaper model needs
What to watch next
- Independent cost-per-task tests that include output tokens and failure rates
- Whether Ember gets an open-weight release, as Spark did with Glimmer
- Whether agent tools such as coding assistants add Ember as a default worker model
- How Anthropic and OpenAI respond on small-model pricing
The practical takeaway is that the future of agents is not one giant model doing everything. It is a team: one expensive planner and a lot of cheap workers. If you build or buy agent software, the question to ask is no longer “which model is best?” but “which model should do which step?” Ember is Meta’s answer for the step where there are a lot of repetitions and not much glory.
Sources and further reading
- TechCrunch: Meta releases Muse Ember, a new lightweight model for agent execution
- Artificial Analysis: Intelligence Index
Featured image: server racks in a data center, free stock photo via Pexels.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
